Multi-algorithm joint feature selection for 28-day prognosis in critically ill patients with colorectal cancer: lightGBM-based model development and multicenter validation

Abstract Objective Colorectal cancer (CRC) is one of the most common malignancies globally and among the elderly population, characterized by a high incidence rate and complex chronic comorbidities. Severe CRC patients experience a high mortality rate after admission to the intensive care unit (ICU). Accurate prognostic prediction is crucial for the optimal allocation of clinical resources and the formulation of personalized treatment decisions. This study aims to develop a machine learning prognostic model based on routine clinical features to systematically evaluate and predict the 28-day all-cause mortality rate in severe elderly CRC patients following ICU admission. Methods This retrospective cohort study extracted data of critically ill patients with CRC from the MIMIC-IV database for model training and internal validation, with the eICU-CRD serving as an independent external validation cohort. The primary outcome was 28-day all-cause mortality following ICU admission. During data preprocessing, variables exhibiting multicollinearity were eliminated using Generalized Variance Inflation Factors (GVIF). To prevent data leakage, robust predictors were extracted using a nested ten-fold cross-validation approach combining LASSO, Random Forest, and the Genetic Algorithm. While maintaining the natural, real-world class distribution, 20 machine learning algorithms were systematically evaluated utilizing a stratified five-times repeated ten-fold cross-validation framework and 1000-iteration Bootstrap resampling. Finally, the optimal model was thoroughly dissected using the SHAP and LIME interpretability frameworks and deployed as an interactive web application to assist in clinical decision-making. Results Through the cross-validation of the three feature selection algorithms, 14 core prognostic features were successfully identified, including the Simplified Acute Physiology Score II (sapsii_score), Red Cell Distribution Width (rdw), charlson_comorbidity_index, and lactate, among others. Of the 20 algorithms evaluated, the LightGBM model demonstrated the most outstanding overall performance. In internal testing, its Area Under the Receiver Operating Characteristic Curve (AUC) achieved the highest overall score of 0.8253 (0.7824, 0.8648), with an accuracy of 0.8434, a specificity of 0.9745, and a Brier Score of 0.1106. Additionally, it exhibited significant net clinical benefit in the Decision Curve Analysis (DCA). However, in the independent external validation set (eICU), which was characterized by a small sample size and extreme class imbalance, the model’s predictive efficacy was significantly attenuated. The AUC dropped to 0.615 (95% CI: 0.413–0.817), revealing notable negative bias and a tendency toward risk overestimation. Conclusion For severe elderly patients with colorectal cancer, the prognostic model constructed based on the LightGBM algorithm demonstrated excellent risk identification accuracy and probability calibration in internal validation. Although the small sample size of the independent external validation cohort posed certain limitations for assessing the model’s generalization, this study introduced the dual frameworks of SHAP and LIME to provide visual interpretations of prognostic features. It successfully developed a Streamlit-based interactive Web application ( https://icu-mortality-risk-predictor-for-crc-cancer-patients.streamlit.app/ ). This enables the prognostic tool not only to effectively achieve early patient risk stratification but also to genuinely empower frontline physicians by providing robust, evidence-based support for optimizing personalized intervention plans and making key clinical diagnostic and treatment decisions.